Function predict

eval/run_predict_os_gensis.py:282–382

(args, datasets)

    return (episode,image_path,history_list,use_low_instruction)

def predict(args, datasets):
    # set global variable
    global USE_LOW_INSTRUCTION
    USE_LOW_INSTRUCTION = (args.data_name == 'android_control_low_test')
    data_dir = args.data_dir
    split_type = args.split
    print("Predicting on:",datasets)
    
    
    if multiprocessing.get_start_method(allow_none=True) != "spawn":
        multiprocessing.set_start_method("spawn", force=True)
    
    with ProcessPoolExecutor(max_workers=len(DEVICES),initializer=_init_llm,initargs=(args.model_path,)) as poolexec:
        tasks = []
        print("Moving model to devices")
        for device in DEVICES:
            tasks.append(poolexec.submit(move_to, device))
        for t in tasks:
            print(t.result())
    
        for dataset in datasets:
            save_dir = os.path.join(args.output_dir, dataset)
            if not os.path.exists(save_dir):
                os.makedirs(save_dir)
                
            episode_dir = os.path.join(data_dir, split_type, dataset)

            # Use predict.jsonl file to store results (write line by line)
            output_file = os.path.join(save_dir, "predict.jsonl")
            
            # Get the list of all episodes files
            if os.path.exists(episode_dir):
                episodes_files = os.listdir(episode_dir)
            else:
                continue
            
            future = []
            all_tasks = []
            print("Loading episodes")
            with ThreadPoolExecutor(max_workers=16) as executor:
                for episodes_file in episodes_files:

                    episodes_path = os.path.join(episode_dir, episodes_file, f"{episodes_file}.json")
                    try:
                        with open(episodes_path, 'r', encoding='utf-8') as f:
                            episodes = json.load(f)
                    except Exception as e:
                        print(f"Failed to load {episodes_path}: {e}")
                        continue
                        # Skip this file on error
                    for index,episode in enumerate(episodes):
                        episode_history = []  # Create a separate history for each episode
                        for prev_episode in episodes[:index]:
                        #for prev_episode in episodes[:episode['step_id']-1]:  # Only get history before current step
                            image_path = os.path.join(episode_dir, episodes_file, f"{episodes_file}_{prev_episode['step_id']}.jpeg")
                            if not os.path.exists(image_path):
                                image_path = image_path.replace(".jpeg", ".png")
                            if not os.path.exists(image_path):